List your company
ResearchFormerly WizWand

Image Classification on Aircraft (Acc, Rob)

32.7Top-1 Clean AccuracyMTA
0.07528.545117.01525.4849May 25, 2026May 29, 2026Jun 3, 2026Jun 8, 2026Jun 12, 2026Jun 17, 2026Jun 22, 2026
Updated 4mo ago

Evaluation results

MethodLinks
MTABackbone=ViT-L/14, Perturbation budget (epsilon)=4/2552026.0632.7——8
T-VSSBackbone=ViT-L/14, Perturbation budget (epsilon)=4/2552026.0632.7——26.9
TTCBackbone=ViT-L/14, Perturbation budget (epsilon)=4/2552026.0631.7——0.5
EnsembleBackbone=ViT-L/14, Perturbation budget (epsilon)=4/2552026.0631.7——12.7
R-TPTBackbone=ViT-L/14, Perturbation budget (epsilon)=4/2552026.0631.7——17.2
CLIPBackbone=ViT-L/14, Perturbation budget (epsilon)=4/2552026.0630——0
TTPBackbone=ViT-L/14, Perturbation budget (epsilon)=4/2552026.0629.2——17.7
MTABackbone=ViT-B/16, Perturbation budget (epsilon)=4/2552026.0625——4.3
T-VSSBackbone=ViT-B/16, Perturbation budget (epsilon)=4/2552026.0624.3——24.4
CLIPBackbone=ViT-B/16, Perturbation budget (epsilon)=4/2552026.0623.9——0
R-TPTBackbone=ViT-B/16, Perturbation budget (epsilon)=4/2552026.0623.9——13.2
TTPBackbone=ViT-B/16, Perturbation budget (epsilon)=4/2552026.0623.9——14.8
AdvMaPLeShots=16, Training Epochs=10, Training Attack (PGD)=2-step (1/255), Evaluation Attack (PGD-100)=100-step (1/255)2026.0523.63——7.33
FAPShots=16, Training Epochs=10, Training Attack (PGD)=2-step (1/255), Evaluation Attack (PGD-100)=100-step (1/255)2026.0523.5——7.97
EnsembleBackbone=ViT-B/16, Perturbation budget (epsilon)=4/2552026.0623.4——8.7
TTCBackbone=ViT-B/16, Perturbation budget (epsilon)=4/2552026.0623.3——0.5
AdvVLPShots=16, Training Epochs=10, Training Attack (PGD)=2-step (1/255), Evaluation Attack (PGD-100)=100-step (1/255)2026.0523.27——8.4
MTABackbone=ResNet-50, Perturbation budget (epsilon)=1/2552026.0618.1——3.7
T-VSSBackbone=ResNet-50, Perturbation budget (epsilon)=1/2552026.0618——20.3
R-TPTBackbone=ResNet-50, Perturbation budget (epsilon)=1/2552026.0617.5——12.6
CLBPShots=16, Training Epochs=10, Training Attack (PGD)=2-step (1/255), Evaluation Attack (PGD-100)=100-step (1/255)2026.0517.46——15.42
TPTBackbone=ResNet-50, Perturbation budget (epsilon)=1/2552026.0617.3——0
C-TPTBackbone=ResNet-50, Perturbation budget (epsilon)=1/2552026.0616.7——0
EnsembleBackbone=ResNet-50, Perturbation budget (epsilon)=1/2552026.0616.4——9.8
CLIPBackbone=ResNet-50, Perturbation budget (epsilon)=1/2552026.0615.7——0
TTPBackbone=ResNet-50, Perturbation budget (epsilon)=1/2552026.0615.6——11.6
AdvVPShots=16, Training Epochs=10, Training Attack (PGD)=2-step (1/255), Evaluation Attack (PGD-100)=100-step (1/255)2026.051.33——0.63
APTBackbone=CLIP-ResNet50, Augmented Views=63, Adversarial perturbation budget (epsilon)=1.02026.06—0.90.9—
APT+TeCoABackbone=CLIP-ResNet50, Augmented Views=63, Adversarial perturbation budget (epsilon)=1.02026.06—9.99.7—
C-TPTSetting=Zero-shot, Backbone=ViT-B/16, Attack=PGD-100, Perturbation Budget (epsilon_a)=4/255, Patch Granularity=16x162026.05—23.90—
CLBPSetting=Zero-shot, Backbone=ViT-B/16, Attack=PGD-100, Perturbation Budget (epsilon_a)=4/255, Patch Granularity=16x162026.05—21.814.6—
CLIPSetting=Zero-shot, Backbone=ViT-B/16, Attack=PGD-100, Perturbation Budget (epsilon_a)=4/255, Patch Granularity=16x162026.05—23.90—
CLIPBackbone=CLIP-ResNet50, Augmented Views=63, Adversarial perturbation budget (epsilon)=1.02026.06—15.70—
MTASetting=Zero-shot, Backbone=ViT-B/16, Attack=PGD-100, Perturbation Budget (epsilon_a)=4/255, Patch Granularity=16x162026.05—254.3—
R-TPTSetting=Zero-shot, Backbone=ViT-B/16, Attack=PGD-100, Perturbation Budget (epsilon_a)=4/255, Patch Granularity=16x162026.05—23.913.2—
SS-TPTBackbone=CLIP-ResNet50, Augmented Views=63, Adversarial perturbation budget (epsilon)=1.0, mode=test-time2026.06—18.213.6—
TeCoABackbone=CLIP-ResNet50, Augmented Views=63, Adversarial perturbation budget (epsilon)=1.02026.06—5.85.8—
TPTSetting=Zero-shot, Backbone=ViT-B/16, Attack=PGD-100, Perturbation Budget (epsilon_a)=4/255, Patch Granularity=16x162026.05—23.40—